A comparison of artificial neural networks and support vector machines for short-term load forecasting using various load types

Abstract

The electric utility depends on accurate load predictions for scheduling spinning reserve, unit commitment, fuel allocation and maintenance. Artificial neural networks (ANNs) have been the popular method for load forecasting but support vector machines (SVM) have recently been successfully applied to the challenge of load forecasting. In this study, the ANN and SVM techniques were applied to short term (one hour) load forecasting on a small island power system of Trinidad and Tobago for three load types. These load types are batch, continuous and batch-continuous load types which represent three unique industrial customers. A performance comparison between the ANN and SVM showed that the SVM produced repeatability, always yielding the global minimum. Both ANN and SVM were unable to accurately perform load forecasts where there may be erratic load patterns or missing data, yielding deviations greater than 3%. For the continuously varying load, the ANN and SVM load forecasts yielded a maximum deviation of 1.20%.